Static facades for using TensorFlow in ScalaPy. Currently contains minimal bindings for usage with TensorFlow.
To install scalapy-tensorflow use JitPack
for scala 2.13:
scalaVersion := "2.13.3"
resolvers in ThisBuild += "jitpack" at "https://jitpack.io"
libraryDependencies += "com.github.VirtuslabRnD.scalapy-tensorflow" %% "scalapy-tensorflow-jvm" % "develop-SNAPSHOT"
for dotty:
scalaVersion := "0.26.0-RC1"
resolvers in ThisBuild += "jitpack" at "https://jitpack.io"
libraryDependencies += "com.github.VirtuslabRnD.scalapy-tensorflow" %% "dotty-tensorflow" % "develop-SNAPSHOT"
Libraries required to run scala-native are listed here: https://scala-native.readthedocs.io/en/v0.3.9-docs/user/setup.html.
Likely just libunwind and re2 should be enough on a typical Linux distro.
To quickly start using this library a development container image with all necessary configuration has been provided Dockerfile.bench.
Use container either via VS Code dev containers or using docker client by running this commands:
docker build -f Dockerfile.bench . -t tensorflow
docker run -v `pwd`:/scalapy-tensorflow/ -w /scalapy-tensorflow/ -it tensorflow /bin/bash
Running performance testing in the aforementioned container comparing BidirectionalLSTMExample.scala with the python version BidirectionalLSTMExample.py showed there is no noticeable difference in learning speed. It's expected as the learning happens on the python side either way.
Results:
| Scala | Python | |
|---|---|---|
| Total time | 108.5 | 107.8 |
| Epoch 1 | 46s 59ms/step | 41s 52ms/step |
| Epoch 2 | 41s 53ms/step | 48s 61ms/step |
| Loss | 0.64 | 0.64 |
| Accuracy | 0.61 | 0.61 |